Comparison of GARCH, LSTM, and Hybrid GARCH-LSTM Models for Analyzing Data Volatility

Abstract

Most statistical methods in time series analytics assume that the residuals are independently and identically distributed with zero mean and constant variance. In real cases, this assumption may be violated. Nowadays, data are dynamic and highly volatile, particularly in finance. The Generalized Autoregressive Conditional Heteroscedasticity (GARCH) model is a statistical method for non-constant conditional variance that can capture the volatility data. Recently, artificial intelligence methods are gaining popularity and have promising performance, one of those is the Long Short-Term Memory (LSTM) method. However, due to the filtering process by forget gate in the LSTM cell some information is missing, which can decrease the prediction’s accuracy. This study proposes a method, namely Hybrid GARCH-LSTM, to overcome those limitation. The performance of the proposed method is evaluated in the simulation and empirical data and compared with GARCH and LSTM model. The results show that the Hybrid GARCH–LSTM model is able to recognize the volatility pattern of data well and outperforms all the other models.